A geometric deep learning approach for checking element-to-entity mappings in infrastructure building information models. Issue 1 (23rd November 2020)

• Record Type: Journal Article
• Authors: Koo, Bonsang; Jung, Raekyu; Yu, Youngsu; Kim, Inhan
• Is Part Of: Journal of computational design and engineering. Volume 8:Issue 1(2021)
• Page Start: 239
• Page End: 250
• Publication Date: 2020-11-23
• Subjects: BIM -- IFC -- geometric deep learning -- semantic integrity -- infrastructure; Engineering -- Data processing -- Periodicals; Computer-aided design -- Periodicals; Computer-aided design; Engineering -- Data processing; Electronic journals; Electronic journals; Periodicals; 620.0042
• DOI: 10.1093/jcde/qwaa075
• Languages: English
• ISSNs: 2288-4300

Abstract

Abstract: Data interoperability between domain-specific applications is a key prerequisite for building information modeling (BIM) to solidify its position as a central medium for collaboration and information sharing in the construction industry. The Industry Foundation Classes (IFC) provides an open and neutral data format to standardize data exchanges in BIM, but is often exposed to data loss and misclassifications. Concretely, errors in mappings between BIM elements and IFC entities may occur due to manual omissions or the lack of awareness of the IFC schema itself, which is broadly defined and highly complex. This study explored the use of geometric deep learning models to classify infrastructure BIM elements, with the ultimate goal of automating the prechecking of BIM-to-IFC mappings. Two models with proven classification performance, Multi-View Convolutional Neural Network (MVCNN) and PointNet, were trained and tested to classify 10 types of commonly used BIM elements in road infrastructure, using a dataset of 1496 3D models. Results revealed MVCNN as the superior model with ACC and F1 score values of 0.98 and 0.98, compared with PointNet's corresponding values of 0.83 and 0.87, respectively. MVCNN, which employs multiple images to learn the features of a 3D artifact, was able to discern subtle differences in their shapes and geometry. PointNet seems to lose the granularity of the shapes, as it uses points partially selected from point clouds. Graphical Abstract: